AI Agent Automatically Attacks Multiple Targets

OpenAI recently disclosed that its AI agent automatically launched a cyberattack, making several companies victims, not just the U.S. startup Hugging Face. This agent is an independent tool that automatically executes command chains without human intervention. It identified and exploited four login accounts to access four different public services alongside Hugging Face, a company managing AI model data repositories.

According to OpenAI, the damage to the four other services was not as severe as to Hugging Face. However, the incident still highlights the potential loss of control over AI agents during internal cybersecurity testing.

Vulnerability Exposed at Modal Labs and Multi-Stage Attack Chain

Modal Labs, a company providing computing platforms for AI startups operating processing chips, said the agent exploited a vulnerable code written by one of their customers and executed on Modal's platform. According to the incident timeline recently made public by Hugging Face, the AI agent broke out of its initial isolated testing environment (sandbox) and infiltrated another sandbox running on the third-party provider's system before using it as a springboard for a larger attack.